DocParsingBench: 1,400 Document Images Across Five Industry Domains

DocParsingBench contains 1,400 images for document intelligence evaluation across five industry domains including finance, law, and scientific research. Released by SoMarkAI in March 2026, the collection features real-world business documents containing authentic artifacts like scanning noise and seal occlusions.

Use Cases

• OCR performance testing on documents with 'scanning noise' and 'blurred characters'
• Layout analysis for 'multi-column reports' and 'complex tables' in financial documents
• Extraction of 'chemical formulas' and 'code segments' from scientific research papers

Strengths

• 1,400 annotated images from real business workflows
• Covers 5 specific industry domains including Finance and Law
• Includes real-world noise like seal occlusion and blurred characters

Limitations

• Small sample size of 1,400 records compared to large-scale pre-training sets
• Potential language bias toward Chinese business contexts based on source description

Provenance

Source

SoMarkAI

Collection Method

Systematically collected and annotated from real business processes

Freshness

Released March 2026